Enzyme fertilizer intelligent ingredient optimization method and system

Abstract

This application relates to the field of bio-fermentation ingredient technology, and discloses an intelligent ingredient optimization method and system for enzyme fertilizer. The method includes: collecting batch information and bioactivity characteristics of raw materials, calculating the carbon-nitrogen ratio balance coefficient, and fusing them to obtain a comprehensive feature vector; calculating the fermentation synergy effect coefficient based on historical ingredient and finished product data, combined with deep correlation features, and constructing a synergy effect matrix; constructing and solving a multi-objective ingredient optimization model to obtain the optimal ingredient scheme; feeding materials according to the scheme and collecting real-time fermentation data, comparing it with the predicted results to obtain the deviation; updating the synergy effect matrix with new sample increments when the deviation exceeds the range; associating finished product quality, ingredient scheme, updated matrix, and batch information to form a structured knowledge record; constructing an ingredient knowledge graph after the record reaches a preset number, and analyzing deep correlation features as prior data to participate in subsequent optimization. This application effectively improves the adaptability of fermentation ingredients and the stability of finished product quality.

Description

Technical Field

[0001] This application relates to the field of bio-fermentation ingredient technology, and in particular to an intelligent ingredient optimization method and system for enzyme fertilizer. Background Technology

[0002] Enzyme fertilizer is a bio-organic fertilizer made from organic waste through fermentation with enzyme strains. It contains active enzymes, beneficial microorganisms, and various nutrients, playing an important role in improving soil structure and promoting crop growth. In the production of enzyme fertilizer, ingredient formulation is a crucial preliminary step that determines the quality of the finished product. The rationality of the ingredient formulation directly affects the efficiency of the subsequent fermentation process, the nutrient compliance rate of the finished product, and the enzyme activity level.

[0003] Enzyme fertilizer raw materials come from a wide range of sources, including livestock and poultry manure, crop straw, fruit pomace, sugar residue, and enzyme strains. The nutrient content, organic matter ratio, enzyme activity, and microbial community state of each batch of raw materials exhibit inherent fluctuations. Current formulation technologies mostly use fixed formulas, failing to systematically collect and vectorize the biological activity characteristics of the raw materials. This makes it impossible to dynamically adjust the formulation based on the activity differences between batches of raw materials, resulting in difficulty in ensuring the consistency of the finished enzyme activity. Furthermore, nonlinear synergistic effects exist when multiple raw materials are fermented together. Current technologies lack quantitative calculation methods for the synergistic effects of different raw material combinations and fail to incorporate deep correlations extracted from historical data into the synergistic effect evaluation process. Formulation decisions remain at the level of linear superposition of static nutrient indicators, making it difficult to comprehensively consider the interactions between raw materials from the perspective of maximizing enzyme activity.

[0004] Therefore, it is urgent to provide an intelligent formulation optimization method and system for enzyme fertilizers to solve the problems of existing technologies lacking systematic characterization methods for the bioactivity characteristics of raw materials, lacking quantitative calculation methods for the synergistic effects of fermentation among raw materials, and the information separation between the formulation and fermentation stages, which prevents historical fermentation patterns from being used as prior knowledge to drive continuous optimization of formulation schemes. Summary of the Invention

[0005] This application provides a method and system for optimizing the intelligent batching of enzyme fertilizer, which is used to improve the adaptability of fermentation ingredients and the stability of finished product quality.

[0006] In a first aspect, this application provides a method for optimizing the intelligent formulation of enzyme fertilizer, including: S1. Collect batch information and characteristic parameters of each batch of raw materials in the batching equipment, extract bioactivity characteristics, calculate the dynamic balance coefficient of carbon-nitrogen ratio, and fuse them to obtain the comprehensive feature vector of each batch of raw materials. S2. Obtain historical ingredient plans and corresponding fermented product data to establish a standardized dataset. Combined with the deep correlation features generated in the previous round, calculate the fermentation synergy coefficient of different raw material combinations and construct a fermentation synergy matrix. S3. Construct a multi-objective ingredient optimization model based on the comprehensive feature vector and the fermentation synergy effect matrix, solve the multi-objective ingredient optimization model, and obtain the optimal ingredient scheme; S4. Obtain the fermentation prediction result corresponding to the optimal ingredient scheme, feed the raw materials according to the optimal ingredient scheme, collect real-time fermentation data during the fermentation process, and compare it with the fermentation prediction result to obtain the fermentation deviation result. S5. When the fermentation deviation results exceed the preset deviation range, record the optimal ingredient scheme and real-time fermentation data. Use the optimal ingredient scheme, real-time fermentation data and finished product quality data as new samples, and use the incremental learning method to update the fermentation synergy matrix. S6. Link the finished product quality data, the optimal ingredient formula, the updated fermentation synergy matrix, and the batch information of raw materials to form a structured ingredient knowledge record; S7. When the ingredient knowledge records reach the preset number, construct an ingredient knowledge graph based on all ingredient knowledge records, analyze the deep correlation characteristics between ingredients and finished product quality, and use the deep correlation characteristics as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of fermentation synergy effect matrix.

[0007] Secondly, this application provides an intelligent ingredient optimization system for enzyme fertilizer, comprising: The feature extraction module is used to collect batch information and feature parameters of each batch of raw materials in the batching equipment, extract bioactivity features, calculate the dynamic balance coefficient of carbon-nitrogen ratio, and fuse them to obtain the comprehensive feature vector of each batch of raw materials. The effect modeling module is used to obtain historical ingredient schemes and corresponding fermented product data to establish a standardized dataset, and combine it with the deep correlation features generated in the previous round to calculate the fermentation synergy effect coefficients of different raw material combinations and construct a fermentation synergy effect matrix. The optimization and solution module is used to construct a multi-objective ingredient optimization model based on the comprehensive feature vector and the fermentation synergy effect matrix, solve the multi-objective ingredient optimization model, and obtain the optimal ingredient scheme. The fermentation comparison module is used to obtain the fermentation prediction results corresponding to the optimal ingredient scheme, feed raw materials according to the optimal ingredient scheme, collect real-time fermentation data during the fermentation process, and compare it with the fermentation prediction results to obtain the fermentation deviation results. The incremental update module is used to record the optimal ingredient scheme and real-time fermentation data when the fermentation deviation results exceed the preset deviation range. The optimal ingredient scheme, real-time fermentation data and finished product quality data are used as new samples, and the fermentation synergy matrix is ​​updated using an incremental learning method. The knowledge recording module is used to associate finished product quality data, optimal ingredient scheme, updated fermentation synergy matrix and raw material batch information to form a structured ingredient knowledge record; The graph construction module is used to construct an ingredient knowledge graph based on all ingredient knowledge records when the preset number of ingredient knowledge records are reached. It analyzes the deep correlation characteristics between ingredients and finished product quality, and uses the deep correlation characteristics as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of fermentation synergy effect matrix.

[0008] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described intelligent batching optimization method for enzyme fertilizer.

[0009] Compared with the prior art, the beneficial effects of this application are at least as follows: 1. By collecting batch information and bioactivity characteristics of raw materials and integrating them into a comprehensive feature vector, the problem of existing technologies not systematically characterizing the bioactivity of raw materials is solved, enabling the formulation scheme to be dynamically adapted according to the differences in the activity of raw materials.

[0010] 2. By combining deep correlation features to calculate the fermentation synergy coefficient and construct the synergy matrix, the problem of the lack of quantitative means for synergistic effects among raw materials in existing technologies is solved, providing a decision basis for multi-objective ingredient optimization that reflects the interaction of biological activities.

[0011] 3. By driving incremental learning through fermentation deviation to update the synergistic effect matrix, the problem of information separation between the ingredient preparation and fermentation stages is solved, and an optimization mechanism from fermentation feedback to self-correction of the ingredient preparation model is established.

[0012] 4. By constructing an ingredient knowledge graph and mining deep correlation features as prior data to participate in the next round of optimization, the problem of historical fermentation patterns being unable to systematically drive ingredient improvement is solved, enabling the ingredient optimization process to have continuous self-learning capabilities. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of the intelligent batching optimization method for enzyme fertilizer in this application; Figure 2 This is a diagram showing the distribution of the dynamic equilibrium coefficient of the carbon-nitrogen ratio in this application; Figure 3 This is a three-dimensional response surface plot of the fermentation synergistic effect in this application; Figure 4 Optimize the three-dimensional scatter plot of the Pareto front for the multi-objective purpose of this application; Figure 5 This is a schematic diagram of the intelligent batching optimization system for enzyme fertilizer of this application. Detailed Implementation

[0015] This application provides an intelligent formulation optimization method and system for enzyme fertilizer. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent batching optimization method for enzyme fertilizer in this application includes: Step S1: Collect batch information and characteristic parameters of each batch of raw materials in the batching equipment, extract bioactivity characteristics, calculate the dynamic balance coefficient of carbon-nitrogen ratio, and fuse them to obtain the comprehensive feature vector of each batch of raw materials.

[0017] In one specific embodiment, the process of performing step S1 may specifically include the following steps: The raw materials in each batch of the batching equipment are divided into organic waste raw materials, enzyme strain raw materials and auxiliary materials. The characteristic parameters corresponding to each type of raw material are extracted and integrated to form a set of characteristic parameters. The bioactivity characteristics of various raw materials are extracted from the set of characteristic parameters. The bioactivity characteristics include the microbial activity index and microbial community diversity index corresponding to each raw material. Based on the current carbon-nitrogen ratio of each raw material and the preset target optimal carbon-nitrogen ratio for fermentation, the dynamic balance coefficient of the carbon-nitrogen ratio of each raw material is calculated. The set of characteristic parameters, bioactivity characteristics and dynamic balance coefficient of carbon-nitrogen ratio are integrated to construct a comprehensive feature vector for each raw material. The comprehensive feature vector includes four feature subspaces: nutrient dimension, enzyme activity dimension, microbial activity dimension and nutrient balance dimension.

[0018] Specifically, after receiving each batch of raw materials, the batching equipment classifies them into three categories according to their source and functional attributes: organic waste raw materials, enzyme inoculum raw materials, and auxiliary materials. Organic waste raw materials refer to pre-treated organic materials such as livestock and poultry manure, crop straw, and fruit pomace; enzyme inoculum raw materials refer to inoculum containing specific fermentation strains; and auxiliary materials refer to additives such as brown sugar, molasses, and soybean meal used to adjust fermentation conditions. For organic waste raw materials, nitrogen content, phosphorus content, potassium content, organic matter content, and moisture content are extracted as characteristic parameters; for enzyme inoculum raw materials, total bacterial count, protease activity, cellulase activity, and amylase activity are extracted as characteristic parameters; and for auxiliary materials, reducing sugar content, amino acid nitrogen, and carbon-nitrogen ratio are extracted as characteristic parameters. The characteristic parameters of each type of raw material are collected according to the raw material batch number as an associated index to form a characteristic parameter set. This set is stored in the form of a structured data table, with each row corresponding to a raw material batch and each column corresponding to a characteristic parameter, achieving unified integration of multi-source heterogeneous data.

[0019] Bioactivity characteristics, including the microbial activity index and the microbial community diversity index, were extracted from the set of characteristic parameters. The microbial activity index was determined using the fluorescein diacetate hydrolysis method. Specifically, the raw material sample was extracted by shaking in phosphate buffer, filtered, and the filtrate was added to a fluorescein diacetate substrate solution. After incubation under dark conditions, the fluorescence intensity was measured using a fluorescence spectrophotometer. The measured fluorescence intensity was compared with a standard curve to calculate the microbial activity index, which reflects the metabolic activity of the microorganisms in the raw material. The microbial community diversity index was obtained using high-throughput sequencing technology. Specifically, DNA was extracted from the raw material sample, and PCR amplification was performed using universal primers for the 16S rRNA gene. The amplified product was purified and sequenced. The obtained sequence data underwent quality filtering, redundancy removal, and clustering operations to calculate the Shannon diversity index as the microbial community diversity index. This index characterizes the species richness and evenness of the raw material microbial community. After obtaining the carbon and nitrogen content data from the characteristic parameter sets of each raw material, the current carbon-to-nitrogen ratio of each raw material was calculated. The current carbon-to-nitrogen ratio is equal to the carbon content value divided by the nitrogen content value. The preset optimal carbon-to-nitrogen ratio for fermentation is a fixed value within the range of 25:1 to 35:1, set according to the metabolic needs of aerobic microorganisms during enzyme fertilizer fermentation. The calculation logic for the dynamic balance coefficient of the carbon-to-nitrogen ratio is as follows: calculate the absolute value of the difference between the current carbon-to-nitrogen ratio and the target optimal carbon-to-nitrogen ratio for fermentation, and take this absolute value along with a preset minimum positive number. The reciprocal of the sum is used as the dynamic equilibrium coefficient for the carbon-nitrogen ratio. Specifically, let the carbon content of the raw material be C, the nitrogen content be N, and the target optimal carbon-nitrogen ratio for fermentation be R. Then the current carbon-nitrogen ratio is C divided by N, and the dynamic equilibrium coefficient for the carbon-nitrogen ratio is... ,in The value is 0.05. This coefficient reaches its maximum value when the current carbon-to-nitrogen ratio equals the target optimal carbon-to-nitrogen ratio for fermentation. This indicates that the ingredient's suitability in terms of nutritional balance has reached its peak.

[0020] The set of characteristic parameters, bioactivity characteristics, and the dynamic balance coefficient of the carbon-nitrogen ratio are fused to construct a comprehensive feature vector for each raw material. This comprehensive feature vector includes four feature subspaces: nutrient dimension, enzyme activity dimension, microbial activity dimension, and nutrient balance dimension. The nutrient dimension includes nitrogen content, phosphorus content, potassium content, organic matter content, and moisture content; the enzyme activity dimension includes protease activity, cellulase activity, and amylase activity; the microbial activity dimension includes the microbial activity index and the microbial community diversity index; and the nutrient balance dimension includes the dynamic balance coefficient of the carbon-nitrogen ratio. Before vector concatenation, each component undergoes Min-Max standardization to eliminate dimensional differences caused by different dimensions and numerical ranges. Specifically, for each component, the maximum and minimum values ​​of that component across all batches of raw materials are calculated. The original value is then subtracted from the minimum value and divided by the difference between the maximum and minimum values ​​to obtain the standardized value. The standardized value range of each component is compressed to the interval between 0 and 1. In the comprehensive feature vector, each component is concatenated using an equal-weighted approach, meaning each component occupies an independent dimension without additional weighting coefficients. This approach is based on the fact that the four feature subspaces represent raw material properties from four independent perspectives: nutrient supply, enzyme activity catalysis, microbial metabolism, and nutrient balance. The contribution of each perspective to ingredient optimization is dynamically adjusted in the subsequent multi-objective optimization model through a synergistic effect matrix, thus preserving the integrity of the original information of each component during the feature vector construction stage. The concatenation order of the components is as follows: nutrient dimension first, enzyme activity dimension second, microbial activity dimension third, and nutrient balance dimension last, forming the comprehensive feature vector. This vector is stored in the ingredient database indexed by the raw material batch number and serves as input data for the subsequent multi-objective ingredient optimization model. For example, after a batch of livestock and poultry manure raw materials undergoes the above treatment, the nutrient dimension component in its comprehensive feature vector reflects the relative nutrient level of the batch of raw materials, the enzyme activity dimension component characterizes the vitality state of the strain, the microbial activity dimension component reflects the strength of biological activity, and the nutrient balance dimension component indicates the degree of carbon-nitrogen balance adaptation. The four types of feature subspaces together constitute a complete characterization of the raw materials from static chemical composition to dynamic biological activity.

[0021] refer to Figure 2 The figure shows the distribution of the dynamic equilibrium coefficient of the carbon-nitrogen ratio, which illustrates the degree of deviation between the current carbon-nitrogen ratio and the target optimal carbon-nitrogen ratio for different batches of raw materials, as well as the corresponding trend of the dynamic equilibrium coefficient. It reflects the nonlinear mapping relationship that the closer the carbon-nitrogen ratio is to the target value, the higher the equilibrium coefficient.

[0022] Step S2: Obtain historical ingredient plans and corresponding fermented product data to establish a standardized dataset. Combine the deep correlation features generated in the previous round to calculate the fermentation synergy coefficient of different raw material combinations and construct a fermentation synergy matrix.

[0023] In one specific embodiment, the process of performing step S2 may specifically include the following steps: Historical ingredient plans and corresponding fermented product data were extracted from the historical production database. Abnormal batch data were removed and standardized to obtain a standardized dataset. Based on the standardized dataset and combined with the deep association features generated in the previous round, all raw materials were combined in pairs. The ratio of the two raw materials in each combination was used as the independent variable and the enzyme activity of the finished product was used as the dependent variable. The nonlinear response surface was obtained by fitting the local weighted regression scatter smoothing method. Based on the nonlinear response surface, the peak value of enzyme activity under the nutrient constraints of the finished product and its corresponding ratio range are determined. The predicted value of enzyme activity when each raw material is used alone is calculated. The ratio of the peak value of enzyme activity to the predicted value of enzyme activity is determined as the fermentation synergy coefficient of the corresponding raw material combination, and all fermentation synergy coefficients are obtained. All fermentation synergy coefficients are symmetricized and sparsely constrained to form a fermentation synergy matrix.

[0024] Specifically, historical ingredient plans and corresponding fermented product data are extracted from the historical production database. The historical ingredient plans record the proportions of each batch of raw materials, while the fermented product data records the enzyme activity, total nitrogen, phosphorus, and potassium content, organic matter content, and fermentation cycle. Abnormal fermentation batches are removed by setting an upper limit threshold for the fermentation cycle and a lower limit threshold for the enzyme activity. If a batch's fermentation cycle exceeds the upper limit threshold or its enzyme activity falls below the lower limit threshold, that batch is considered abnormal and is removed. The retained data undergoes standardization by applying Z-score standardization to the enzyme activity and calculating the mean activity of all batches. with standard deviation Subtract the original enzyme activity value from each batch. Divide by Standardized enzyme activity values ​​were obtained. The proportions were standardized using Min-Max normalization, where the minimum proportion of each ingredient was subtracted from the maximum proportion, and then divided by the difference between the minimum and maximum proportions, mapping all proportions to the 0-1 range. This process yielded a standardized dataset.

[0025] Based on a standardized dataset and combined with the deep association features generated in the previous round, all pairwise combinations of raw materials are traversed. The deep association features are derived from the graph neural network embedding vectors of the ingredient knowledge graph, representing the potential interaction patterns between raw materials after deep association mining. In the initialization phase, if features from the previous round are not present, they are filled with zero vectors. This strategy aims to avoid introducing prior bias during the cold start phase. For each pairwise combination of raw materials, the ratio of the two raw materials in the combination is used as the independent variable vector X, and the finished enzyme activity is used as the dependent variable vector Y. A local weighted regression scatter smoothing method is employed to fit the nonlinear response surface. Specifically, a cubic polynomial is selected as the local fitting function, and the neighborhood window width is set to cover 40% to 60% of the total sample size. This range is chosen to balance local fitting accuracy with global trend capture. The distance weight uses a cubic weight function, meaning that samples closer to the target fitting point receive higher weights. The weight coefficient is inversely proportional to the cube of the distance. This function is chosen because it assigns higher weights to nearby samples while smoothly reducing the weights of distant samples, consistent with the physical intuition that the interaction of raw materials decreases as the ratio deviates during fermentation. For each fitted point, weighted least squares regression is performed within the neighborhood window. The regression coefficients are iteratively calculated until the coefficient change between two adjacent iterations is less than 0.001. After convergence, the nonlinear response surface of the raw material combination is obtained.

[0026] Based on the nonlinear response surface, the peak enzyme activity and its corresponding ratio range under the nutrient constraints of the finished product are determined. Specifically, under the constraints that the total nitrogen, phosphorus, and potassium content of the finished product is not lower than a preset nutrient threshold and the organic matter content is within the target range, a grid search method is used to discretize the ratio of the two raw materials within the [0,1] interval into several grid points with a preset step size. The predicted enzyme activity value corresponding to each grid point is calculated, and the maximum value satisfying the constraints is extracted as the peak enzyme activity. The ratio range of the two raw materials corresponding to this peak value is recorded. The predicted enzyme activity value when each raw material is used alone is calculated. The ratio of raw material A is set to 1 and the ratio of raw material B is set to 0. Substituting these values ​​into the nonlinear response surface equation, the predicted enzyme activity value when raw material A is used alone is obtained. Similarly, the predicted enzyme activity value when raw material B is used alone is obtained. The arithmetic mean of the two values ​​is used as the baseline predicted activity value for the raw material combination. The peak enzyme activity value is divided by the baseline predicted activity value to obtain the fermentation synergy coefficient of the corresponding raw material combination. After traversing all raw material combinations, all fermentation synergy coefficients are obtained.

[0027] All fermentation synergy coefficients were symmetricized. The arithmetic mean of the synergy coefficients between raw material A and raw material B, and between raw material B and raw material A, was taken as the synergy coefficient for the undirected raw material pair. Subsequently, sparsity constraints were applied, with a sparsity threshold of 0.05. This threshold was determined based on the distribution of the synergy coefficients after Min-Max standardization, ranging from 0.03 to 0.08. It was selected using cross-validation with the goal of minimizing the prediction error of the finished enzyme activity. If the absolute value of the synergy coefficient of a raw material pair was less than the sparsity threshold, the coefficient was set to zero, retaining only raw material combinations with clear synergistic or antagonistic effects. After symmetry and sparsity constraints, all non-zero synergy coefficients were arranged according to the raw material number to form a fermentation synergy matrix. The rows and columns of this matrix corresponded to the raw material numbers, and the matrix elements represented the strength of the fermentation synergy between the corresponding raw materials.

[0028] refer to Figure 3 The figure is a three-dimensional response surface plot of the fermentation synergy effect, which shows the change surface of the enzyme activity prediction value under different ratio combinations of two raw materials. It marks the optimal ratio range corresponding to the peak of the synergy effect and the activity prediction benchmark point when the raw materials are used alone.

[0029] Step S3: Construct a multi-objective ingredient optimization model based on the comprehensive feature vector and the fermentation synergy effect matrix, solve the multi-objective ingredient optimization model, and obtain the optimal ingredient scheme.

[0030] In one specific embodiment, the process of performing step S3 may specifically include the following steps: Based on the comprehensive feature vector and fermentation synergy matrix, three optimization objectives are set: maximizing enzyme activity, minimizing raw material cost, and minimizing fermentation cycle. At the same time, hard constraints on finished product nutrient indicators and soft constraints on carbon-nitrogen ratio balance are set to construct a multi-objective ingredient optimization model. A pre-defined non-dominated sorting genetic algorithm is used to solve the multi-objective batching optimization model. Population screening is performed through fast non-dominated sorting and crowding calculation. After iterative convergence, the Pareto optimal solution set is obtained. The optimal ingredient ratio is selected from the Pareto optimal solution set, based on the preset decision rules.

[0031] Specifically, the multi-objective ingredient optimization model uses the proportions of each ingredient as decision variables and sets three types of objective functions: maximizing enzyme activity, minimizing raw material costs, and minimizing the fermentation cycle. The enzyme activity maximization objective is formed by weighting the proportions of each ingredient with the synergistic effect coefficients in the fermentation synergistic effect matrix, and then adding the contributions of the enzyme activity dimension and the microbial activity dimension from the comprehensive feature vector. A higher function value indicates higher enzyme activity in the fermented product. The raw material cost minimization objective is constructed by linearly weighting the market price of each ingredient with its proportion, directly reflecting the economic cost of the ingredient formulation. The goal of minimizing the fermentation cycle is based on historical data showing the correlation between raw material ratios and fermentation cycles. A cycle prediction model is established using support vector regression. This model takes the raw material ratio vector as input and the fermentation cycle as output, employing a radial basis function (RBF) kernel with a penalty coefficient of 10 and a kernel function parameter of 0.5. This parameter setting is based on the fact that the RBF kernel can effectively capture the nonlinear mapping between the ratios and the cycle. The penalty coefficient of 10 strikes a balance between model complexity and fitting accuracy, while the kernel function parameter of 0.5 controls the influence range of a single training sample, allowing the model to capture nonlinear relationships while avoiding overfitting. Simultaneously, hard constraints on finished product nutrient indicators and soft constraints on carbon-nitrogen ratio balance are embedded. The hard constraints on finished product nutrient indicators are calculated by linearly summing the nutrient dimension components in the comprehensive feature vector with the ratios, requiring the total nitrogen, phosphorus, and potassium content, organic matter content, and moisture content of the finished product to meet the lower limit of national standards. Any scheme that violates this constraint is directly eliminated. The soft constraint on the carbon-nitrogen ratio balance is achieved by calculating the actual carbon-nitrogen ratio of the overall formulation by weighting the carbon and nitrogen contents of each raw material according to their respective proportions. That is, the actual carbon-nitrogen ratio of the overall formulation = Σ(carbon content of each raw material × proportion of that raw material) / Σ(nitrogen content of each raw material × proportion of that raw material). This actual carbon-nitrogen ratio is required to be within ±10% of the target optimal carbon-nitrogen ratio for fermentation. It is important to note that the carbon-nitrogen ratio here uses the original measured values ​​of each raw material, rather than the dynamic balance coefficient of the carbon-nitrogen ratio in step S1, to ensure that the constraint has clear physical interpretability and allows for slight deviations in exchange for improvement space in other objectives. The above three types of objective functions and two types of constraints are combined to form a multi-objective ingredient optimization model.

[0032] A non-dominated sorting genetic algorithm is used to solve a multi-objective ingredient optimization model. The algorithm uses the ingredient ratio as the decision variable and initializes the population using real-number encoding. Each individual corresponds to a set of ingredient ratio schemes, and the population size is set to 150. This value is chosen to balance the breadth of the search space coverage with the computational cost per iteration; a smaller population size may lead to premature convergence to a local optimum, while a larger population size increases computational overhead. The crossover probability is set to 0.85, and the mutation probability is set to 0.08. This parameter combination balances global exploration and local exploitation capabilities. A higher crossover probability promotes information exchange between individuals, while a lower mutation probability maintains population diversity while avoiding excessive random perturbation. In each generation of evolution, fast non-dominated sorting is performed, dividing individuals into multiple non-dominated levels. The smaller the level number, the better the overall optimization performance of the individuals within that level. Crowding is calculated for individuals within the same level, equal to the sum of the normalized distances of their neighbors in each target dimension. Boundary individuals have their crowding set to infinity. This calculation method prioritizes retaining extreme solutions at the target space boundaries and intermediate solutions in sparsely distributed regions, thus maintaining uniform coverage of the Pareto front. Individuals are selected from the parent population using a tournament selection mechanism. Offspring are generated through simulated binary crossover and polynomial mutation. Parent and offspring are merged, and fast non-dominated sorting and crowding calculation are re-executed to select a new generation population of equal size to the parent population. Iteration continues, and convergence is determined when the Pareto front change rate is less than 0.5% for 20 consecutive generations. This condition is based on the premise that further iterations after the Pareto front stabilizes are unlikely to yield significant improvement.

[0033] After iterative convergence, a Pareto optimal solution set is obtained. This set contains multiple non-dominated solutions where the three objectives of enzyme activity, cost, and fermentation cycle cannot be improved simultaneously. The optimal ingredient formulation is selected from the Pareto optimal solution set based on a pre-defined decision rule. The pre-defined decision rule uses the weighted ideal point method, assigning a weight of 0.5 to the objective of maximizing enzyme activity, 0.3 to the objective of minimizing raw material cost, and 0.2 to the objective of minimizing fermentation cycle. This weighting is based on the fact that enzyme activity is the core indicator of finished product quality, hence its highest weight; raw material cost is given the second highest weight due to economic considerations; and fermentation cycle is given a lower weight due to production efficiency. Three ideal points are identified in the Pareto front for maximum enzyme activity, minimum cost, and shortest fermentation cycle. The normalized Euclidean distance from each non-dominated solution to the three ideal points is calculated and weighted, and the solution with the smallest weighted distance is selected as the optimal ingredient formulation. For example, the Pareto optimal solution set obtained after optimizing a batch of ingredients contains 8 non-dominated solutions. Among them, the predicted enzyme activity of a certain solution is 245 U / g, the raw material cost is 680 yuan per ton, and the fermentation cycle is 18 days. This solution has the smallest weighted distance, so it is selected as the optimal ingredient scheme, and the ratio instructions of each raw material are output.

[0034] refer to Figure 4 The figure is a three-dimensional scatter plot of the Pareto front for multi-objective optimization, showing the distribution of the Pareto optimal solution set obtained by the non-dominated sorting genetic algorithm in the three objective spaces of enzyme activity, raw material cost, and fermentation cycle. It also marks the optimal ingredient scheme and ideal reference point position selected by the weighted ideal point method.

[0035] Step S4: Obtain the fermentation prediction result corresponding to the optimal ingredient scheme, feed the raw materials according to the optimal ingredient scheme, collect real-time fermentation data during the fermentation process, and compare it with the fermentation prediction result to obtain the fermentation deviation result.

[0036] In one specific embodiment, the process of performing step S4 may specifically include the following steps: Based on the optimal ingredient ratio, the feeding control instructions for each raw material are generated, and the ingredient feeding execution mechanism is controlled to complete the feeding of raw materials and start the fermentation process. A fermentation process prediction model is established based on the optimal ingredient scheme to predict the changes in temperature, pH value, dissolved oxygen concentration, carbon dioxide release rate and enzyme activity parameters during fermentation, and to generate fermentation prediction results. During the fermentation process, the temperature, pH value, dissolved oxygen concentration, carbon dioxide release rate, and enzyme activity parameters inside the fermenter are collected to obtain real-time fermentation data. The real-time fermentation data is compared one by one with the corresponding parameters in the fermentation prediction results, the deviation values ​​of each parameter are calculated, and the results are integrated to form the fermentation deviation results.

[0037] Specifically, feeding control instructions for each raw material are generated based on the optimal batching scheme, controlling the batching actuator to complete the feeding of raw materials and start the fermentation process. The optimal batching scheme includes the proportion of each raw material. The feeding control instructions convert the proportion into the speed of the screw feeder or the opening degree and duration of the pneumatic valve in each raw material silo. The control instructions are generated based on the flow coefficient calibrated by the batching equipment. The flow coefficient represents the output weight of each raw material silo per unit time. The target feeding weight of each raw material is calculated by multiplying the proportion by the target total feeding amount, and then dividing by the flow coefficient to obtain the feeder running time or valve opening time. After receiving the instructions, the batching actuator executes the feeding in sequence. After each raw material is weighed and verified, it enters the mixing process. After being mixed evenly, it is transferred to the fermentation tank and the fermentation process is started.

[0038] After feeding, a fermentation process prediction model is established based on the optimal ingredient ratio. This model takes the proportions of each raw material in the optimal ratio, and the nutrient and microbial activity components of the comprehensive feature vector as inputs. It outputs temperature, pH, dissolved oxygen concentration, carbon dioxide release rate, and enzyme activity parameters during fermentation, constructing a multi-output regression prediction model. A Long Short-Term Memory (LSTM) network is used as the model architecture. The static input vector is used as repeated input at each time step, expanding it into a temporal input sequence with a dimension equal to [number of time steps, number of raw material types, and the sum of the dimensions of the comprehensive feature vector]. The hidden layers are set to two layers with 64 neurons each. This number of layers and neurons is chosen because two hidden layers are sufficient to capture the temporal dynamics of the fermentation process, and 64 neurons strike a balance between model expressive power and computational cost. The output layer has a dimension of 5, corresponding to temperature, pH, dissolved oxygen concentration, carbon dioxide release rate, and enzyme activity parameters. During training, multiple complete fermentation time-series data sets were extracted from a historical fermentation database. The first 80% was used as the training set, and the last 20% as the validation set. The Adam optimizer was used with a learning rate of 0.001 and a batch size of 32. The model underwent 200 iterations. This number of iterations was chosen because the length of fermentation time-series data is typically between 15 and 25 days, and 200 iterations are sufficient for the model to converge to a stable state. After training, the current optimal ingredient ratio was input into the model, and the changes in various parameters during future fermentation were predicted point-by-point with a time step of 1 hour, generating fermentation prediction results. It should be noted that the input layer dimension is fixed during system initialization based on the total number of all possible raw material types involved in the batching and the comprehensive feature vector of each raw material type. The comprehensive feature vector of each raw material type is constructed using the union of all components of the four feature subspaces in S1. Raw material types lacking corresponding feature components are filled with 0 values ​​to the unified dimension. For raw materials not used in the current optimal batching scheme, their proportions are filled with 0 values. The dimensions corresponding to missing raw materials in the comprehensive feature vector are also filled with 0 values, thus ensuring that the input dimension is strictly consistent with the dimension during model training for each prediction. For example, after inputting the current batch batch proportion data and the comprehensive feature vector, the model predicts that the temperature on day 3 of fermentation will be 45 degrees Celsius, the pH value will be 6.8, and the dissolved oxygen concentration will be 2.5. The carbon dioxide release rate is 0.8 per hour. The enzyme activity is 85%. .

[0039] After fermentation begins, during the fermentation process, temperature sensors, pH electrodes, dissolved oxygen probes, carbon dioxide concentration sensors, and an online enzyme activity monitoring device are installed inside the fermenter. Temperature, pH value, dissolved oxygen concentration, carbon dioxide release rate, and enzyme activity parameters are collected at preset sampling intervals to obtain real-time fermentation data. The sampling interval is set to 10 minutes, a setting designed to capture rapid dynamic changes in the fermentation process while avoiding data redundancy and sensor wear caused by high-frequency sampling. The real-time fermentation data is compared one by one with the corresponding parameters in the fermentation prediction results, aligned by timestamps, to calculate the deviation value of each parameter. The deviation value calculation logic is as follows: subtract the predicted value of the parameter at the corresponding time point from the parameter value in the real-time fermentation data, and then divide by the predicted value to obtain the normalized deviation value. The normalized deviation values ​​of each parameter are integrated in time series to form the fermentation deviation result. The fermentation deviation result is stored in the form of a time series vector, with each time point corresponding to a vector containing five deviation components. This vector represents the degree of deviation of the current fermentation state from the predicted trajectory.

[0040] Step S5: When the fermentation deviation result exceeds the preset deviation range, record the optimal ingredient scheme and real-time fermentation data. Use the optimal ingredient scheme, real-time fermentation data and finished product quality data as new samples, and update the fermentation synergy matrix using an incremental learning method.

[0041] In one specific embodiment, the process of performing step S5 may specifically include the following steps: The fermentation deviation results are compared with the preset deviation range to determine whether the fermentation deviation results exceed the preset deviation range. If so, the sample collection and matrix update process is started to record the optimal ingredient scheme and real-time fermentation data. After the corresponding batch of fermentation is completed, the finished product quality data is obtained, and the optimal ingredient scheme, real-time fermentation data and finished product quality data after fermentation are determined as new samples. An incremental learning method is used to locally correct the raw material combination coefficients of the deviation correlation in the fermentation synergy matrix based on the new samples, so as to obtain the updated fermentation synergy matrix.

[0042] Specifically, the fermentation deviation results are compared with preset deviation ranges. These preset deviation ranges are set for temperature, pH, dissolved oxygen concentration, carbon dioxide release rate, and enzyme activity parameters. The temperature deviation threshold is set at ±15%, the pH deviation threshold at ±0.5%, the dissolved oxygen concentration deviation threshold at ±20%, the carbon dioxide release rate deviation threshold at ±25%, and the enzyme activity deviation threshold at ±18%. These thresholds are set based on the following: a temperature deviation exceeding 15% indicates a shift in fermentation heat balance, possibly caused by abnormal raw material moisture content or abnormal microbial metabolism; a pH deviation exceeding 0.5 indicates an imbalance in the acid-base buffer system, directly affecting enzyme activity; a dissolved oxygen concentration deviation exceeding 20% ​​reflects unsatisfactory oxygen supply conditions; a carbon dioxide release rate deviation exceeding 25% indicates abnormal microbial respiration intensity; and an enzyme activity deviation exceeding 18% indicates a risk of deviation in finished product quality. The normalized deviation value of each parameter in the fermentation deviation results is compared one by one with its corresponding threshold. If the absolute value of the normalized deviation value of any parameter exceeds the corresponding threshold, the fermentation deviation result is determined to be outside the preset deviation range. Once the exceedance is detected, the sample collection and matrix update process is initiated. The proportions of each ingredient in the optimal formulation are recorded, along with the trigger time and real-time fermentation data from previous sampling cycles. After the batch fermentation is complete, finished product quality data is obtained through finished product inspection. This data includes the total nitrogen, phosphorus, and potassium content, organic matter content, enzyme activity, and total microbial colony count. The optimal formulation, real-time fermentation data, and the batch's finished product quality data are identified as new samples. These new samples are linked and stored using the fermentation batch number as an index, forming a complete data record containing the formulation, process trajectory, and final quality state.

[0043] An incremental learning method is employed to locally correct the coefficients of raw material combinations correlated with deviations in the fermentation synergy matrix based on newly added samples. The incremental learning method uses an online update strategy based on stochastic gradient descent. This strategy is based on the principle of correcting only the local coefficients correlated with deviations in the fermentation synergy matrix, avoiding the forgetting of historical knowledge caused by retraining the entire matrix. Specifically, the method involves extracting the proportions of each raw material and the measured values ​​of the finished enzyme activity from the newly added samples. The relative deviation rate between the measured values ​​and the enzyme activity values ​​predicted based on the current synergy matrix is ​​calculated as (measured value - predicted value) / predicted value, yielding a dimensionless residual signal. This residual is then allocated according to the contribution ratio of each raw material combination in the formulation scheme. The contribution ratio is defined as the proportion of the product of the proportions of two raw material combinations to the sum of the products of the proportions of all raw material combinations. The synergy coefficients of each raw material combination are updated using regularized gradient descent. The update magnitude is the residual multiplied by the contribution ratio of that combination, then multiplied by a preset learning rate of 0.01, ranging from 0.005 to 0.05, determined based on the balance between the historical update magnitude and convergence speed of the synergy coefficients. This allocation method avoids attributing fermentation deviations solely to a single raw material combination, reflecting the nonlinear characteristics of multi-raw material synergy. After the update, the corrected fermentation synergy matrix is ​​symmetricized to ensure that the synergy coefficients of raw material A and raw material B remain symmetrical and consistent, resulting in the updated fermentation synergy matrix. This updated fermentation synergy matrix participates in the construction of a multi-objective batching optimization model during the next batch batching, achieving feedback between the batching scheme and the quality of the fermented product.

[0044] Step S6: Link the finished product quality data, the optimal ingredient scheme, the updated fermentation synergy matrix, and the batch information of the raw materials to form a structured ingredient knowledge record.

[0045] In one specific embodiment, the process of performing step S6 may specifically include the following steps: Based on the finished product quality data corresponding to the current fermentation batch, the optimal ingredient scheme, the updated fermentation synergy matrix, and the batch information of the raw materials, determine the multi-source data to be associated; Based on the multi-source data to be associated, the batch information of raw materials is used as a unified association identifier. The finished product quality data, the optimal ingredient scheme and the updated fermentation synergy matrix are matched and bound with the corresponding batch information one by one. The matched full data is then structured and regularized to generate the corresponding batch of structured ingredient knowledge records.

[0046] Specifically, after the current fermentation batch is completed, the finished product quality data, optimal ingredient ratio, updated fermentation synergy matrix, and batch information of raw materials are extracted from that batch to determine the multi-source data to be correlated. Finished product quality data includes the total nitrogen, phosphorus, and potassium content, organic matter content, enzyme activity, and total microbial colony count. These data are obtained through laboratory testing after fermentation and reflect the actual performance of the batch's ingredient ratio at the fermentation terminal. The optimal ingredient ratio records the specific proportions of each raw material in the batch and is the output of the ingredient optimization model. The updated fermentation synergy matrix contains the synergy coefficients between raw materials after incremental learning correction during the batch's fermentation process, representing the corrections made to the original model by the batch's fermentation experience. The batch information of raw materials refers to the unique number assigned to each raw material upon entry into the warehouse. This number is generated when the raw material enters the ingredient mixing equipment and is bound to the raw material's comprehensive feature vector, persisting throughout the entire process of ingredient mixing, fermentation, and finished product inspection. These four types of data, with different sources and formats, together constitute the multi-source data to be correlated.

[0047] Based on the multi-source data to be associated, batch information of raw materials is used as a unified association identifier. Finished product quality data, optimal ingredient formulation, and the updated fermentation synergy matrix are matched and bound one-to-one with the corresponding batch information. The basis for selecting raw material batch information as the unified association identifier is that batch information is a unique identifier throughout the entire chain of raw material procurement, ingredient formulation, fermentation, and finished product inspection. Using this identifier as the association key ensures that multi-source data points to the same physical batch, avoiding cross-batch data confusion. The matching and binding process is as follows: the batch number of each raw material in the batch is retrieved from the ingredient database; the corresponding finished product quality data is extracted from the finished product inspection database based on the batch number; the optimal ingredient formulation for the batch is extracted from the ingredient formulation storage area; and the updated fermentation synergy matrix for the batch is extracted from the model parameter storage area. These four types of data are aligned according to the batch number to form a binding relationship. The matched full data undergoes structured processing, including removing outlier null values, standardizing numerical precision, standardizing timestamp formats, and verifying data integrity, transforming heterogeneous multi-source data into structured data in a unified format. For example, a batch of raw materials with batch number F20250708A, after matching and binding, the finished product quality data associated with this batch number is 5.6% total nitrogen, phosphorus and potassium, 48% organic matter, and 238 U / g enzyme activity. In the optimal ingredient scheme, livestock and poultry manure accounts for 42% and crop straw accounts for 18%. In the updated fermentation synergy matrix, the combination coefficient of livestock and poultry manure and fruit pomace is corrected from 1.12 to 1.08. After the above data is structured and normalized, a structured ingredient knowledge record for the corresponding batch is generated. This record is stored in the form of a relational data table and contains the entire chain information of the batch from raw material attributes, ingredient decision, fermentation correction to finished product quality.

[0048] Step S7: When the ingredient knowledge records reach the preset number, construct an ingredient knowledge graph based on all ingredient knowledge records, analyze the deep correlation characteristics between ingredients and finished product quality, and use the deep correlation characteristics as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of fermentation synergy effect matrix.

[0049] In one specific embodiment, the process of performing step S7 may specifically include the following steps: Count the total number of generated ingredient knowledge records. When the total number reaches the preset number, retrieve all ingredient knowledge records as the data source for graph construction. Based on the graph, a data source is constructed, and the correspondence between the ingredient parameters and the finished product quality is extracted from each ingredient knowledge record to form an ingredient knowledge graph. By performing association mining analysis on the ingredient knowledge graph, we can extract the deep correlation features between ingredients and finished product quality. Deeply correlated features are used as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of the fermentation synergy effect matrix.

[0050] Specifically, the total number of generated ingredient knowledge records is counted. These records are stored in the ingredient database as structured tables, with each record containing raw material batch information, the optimal ingredient formulation, the updated fermentation synergy matrix, and finished product quality data. The ingredient database is traversed, and records are counted based on their generation timestamps, accumulating to obtain the total number. A preset number of 500 records is set. This value is based on the following: too few records make it difficult for the graph neural network to learn stable association patterns, while too many records lead to an excessively long graph construction cycle, failing to respond promptly to seasonal changes in raw material characteristics. 500 records typically correspond to 6 to 8 months of production accumulation, covering the main cycles of raw material batch fluctuations. When the total number reaches the preset number, the graph construction process is triggered, retrieving all ingredient knowledge records as the graph construction data source. The retrieval process performs a full read based on the record generation timestamps, loading the scattered structured ingredient knowledge records into the memory cache to form the graph construction data source. When the total number does not reach the preset number, the deep association feature is initialized with a zero vector to ensure that the variable can still be obtained in step S2 before the ingredient knowledge graph is constructed, so as to form a complete computational closed loop.

[0051] Based on the knowledge graph, a data source is constructed to extract the correspondence between ingredient parameters and finished product quality from each ingredient knowledge record. Ingredient parameters include the type of each raw material, the ratio, the dynamic balance coefficient of the carbon-nitrogen ratio, and the nutrient dimension component in the comprehensive feature vector. Finished product quality includes the total nitrogen, phosphorus, and potassium content, organic matter content, enzyme activity, and total microbial colony count. When constructing the ingredient knowledge graph, a heterogeneous graph structure is adopted, with raw material type, ratio range, and dynamic balance coefficient range of carbon-nitrogen ratio as different types of nodes, and finished product quality indicators as independent quality node types added to the graph. Each type of node has an independent feature space and encoding method, with the correlation strength between ingredient parameters and finished product quality indicators as the edges. Specifically, the ratio range is discretized into three intervals—low (0 to 33% quantile), medium (33% to 67% quantile), and high (67% quantile to 100%)—based on the equal frequency discretization of the ratio distribution of each raw material in all historical records; the dynamic balance coefficient range of the carbon-nitrogen ratio is uniformly divided from 0 to the maximum value with a step size of 0.1, and each interval spans 0.1. Specifically, the implementation involves mapping each raw material type to a raw material node in the graph, discretizing the proportions into three intervals (low, medium, and high) and mapping them to proportion nodes, and mapping the finished product quality indicator compliance status to quality nodes. Edge relationships include raw material co-occurrence relationships, causal relationships between proportions and quality, and temporal relationships between carbon-nitrogen ratio and fermentation cycle. The weight of each edge is calculated by combining the frequency of occurrence of the corresponding relationship in historical records and the finished product quality compliance rate. For example, in a certain ingredient knowledge record, if the proportion of livestock and poultry manure is 42%, crop straw is 18%, and the finished product enzyme activity is 238 U / g, then a co-occurrence edge is established between the livestock and poultry manure node and the 42% proportion interval node in the graph, and a causal edge is established between this proportion interval node and the enzyme activity compliance status node. When a new raw material type not included in the historical records appears, a corresponding raw material node is added to the graph. The initial feature vector of this new node is generated as follows: nutrient dimension, enzyme activity dimension, and microbial activity dimension components are extracted from the comprehensive feature vector of the raw material, concatenated, and mapped to the same feature space as existing raw material nodes through a linear transformation matrix. In each graph reconstruction, the linear transformation matrix uses the original features of the existing raw material nodes as independent variables and their embedded vectors trained by the graph neural network as dependent variables. Through least squares fitting, the transformed new node features are aligned with the existing node embedding space, thus ensuring that the new nodes can be seamlessly integrated into the existing graph structure. The number of columns in the graph neural network input feature matrix remains consistent.

[0052] Association mining analysis was performed on the ingredient knowledge graph to extract deep correlation features between ingredients and finished product quality. The association mining analysis employed a graph neural network for embedding learning. The graph neural network consisted of two graph convolutional layers and one fully connected output layer. The graph convolutional layer dimension was set to 128, a value chosen because 128-dimensional embedding vectors can adequately represent the high-order neighborhood information of nodes in the ingredient knowledge graph while keeping computational complexity within an acceptable range. During training, the node feature matrix and adjacency matrix of the ingredient knowledge graph were input into the graph neural network. The node feature matrix was formed by concatenating ingredient parameters and finished product quality indicators, and the adjacency matrix consisted of the association strength of each edge in the graph. A negative sampling strategy was used for training, with a negative sampling ratio of 1:5, meaning that for every positive sample edge, five negative sample edges were generated. This ratio was chosen because effective association edges in the ingredient knowledge graph are relatively sparse, and negative sampling enhances the model's ability to distinguish invalid associations. The training iterations were 100 rounds, with a learning rate set to 0.005. This learning rate was chosen because graph neural networks converge slowly on knowledge graphs, and a lower learning rate helps to stably capture deep associations. Through forward propagation of the graph convolutional layers, each node aggregates the node features within its second-order neighborhood to generate a node embedding vector. Average pooling is then performed on the node embedding vectors to obtain a graph-level embedding vector, which represents the deep association features between ingredients and finished product quality. These deep association features characterize implicit patterns that go beyond the linear superposition of single ingredient parameters. For example, the combination of ingredient type A1 and ingredient type B1 within a specific carbon-nitrogen ratio range has a non-linear promoting effect on the enzyme activity of the finished product; this pattern cannot be directly discovered using traditional statistical methods.

[0053] Deep correlation features are used as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of the fermentation synergy matrix. The prior data is stored in the form of a 128-dimensional embedding vector. When constructing the next round of multi-objective ingredient optimization model, this vector is concatenated to the end of the comprehensive feature vector to form an extended input vector, enabling the optimization model to make ingredient decisions based on historical deep correlation patterns. When generating the fermentation synergy matrix, the prior data is used to initialize the unknown raw material combination coefficients in the matrix. Specifically, the cosine similarity of the embedding vectors of each raw material node in the deep correlation features is calculated, and the similarity value is linearly scaled, i.e., (similarity + 1) / 2, and mapped from the interval [-1, 1] to the interval [0, 1] as the initialization weight. This weight is assigned to the raw material combinations in the fermentation synergy matrix that lack historical data. The basis of this initialization strategy is that raw materials that are closely related in the ingredient knowledge graph usually have a positive correlation in their fermentation synergy effects, and cosine similarity can effectively quantify this correlation. In the next batch of ingredients, the comprehensive feature vector of the new batch of raw materials and the expanded deep correlation features are jointly input into the multi-objective ingredient optimization model. The fermentation synergy matrix carries historical deep correlation information during the initialization stage, thereby improving the model's generalization ability when facing new raw materials or new formulation requirements. For example, when a new type of fruit pomace raw material is introduced, the synergy coefficient between this raw material and existing raw materials in the fermentation synergy matrix is ​​initialized using the cosine similarity in the deep correlation features, avoiding prediction bias caused by relying entirely on random initialization during the cold start stage.

[0054] The construction of the ingredient knowledge graph and the embedding learning of the graph neural network transform discrete historical ingredient experience into structured knowledge representation. Through the extraction and prior injection of deep correlation features, the multi-objective ingredient optimization model can make decisions by referring to implicit knowledge that transcends explicit statistical laws. The fermentation synergy effect matrix obtains a reasonable initial value when facing new raw material combinations, gradually eliminating model bias caused by changes in raw material sources or seasonal fluctuations. This realizes the closed-loop optimization of the ingredient system from experience accumulation to knowledge reuse.

[0055] The above describes the intelligent batching optimization method for enzyme fertilizer in the embodiments of this application. The following describes the intelligent batching optimization system for enzyme fertilizer in the embodiments of this application. Please refer to [link / reference]. Figure 5 One embodiment of the intelligent batching optimization system for enzyme fertilizer in this application includes: The feature extraction module is used to collect batch information and feature parameters of each batch of raw materials in the batching equipment, extract bioactivity features, calculate the dynamic balance coefficient of carbon-nitrogen ratio, and fuse them to obtain the comprehensive feature vector of each batch of raw materials. The effect modeling module is used to obtain historical ingredient schemes and corresponding fermented product data to establish a standardized dataset, and combine it with the deep correlation features generated in the previous round to calculate the fermentation synergy effect coefficients of different raw material combinations and construct a fermentation synergy effect matrix. The optimization and solution module is used to construct a multi-objective ingredient optimization model based on the comprehensive feature vector and the fermentation synergy effect matrix, solve the multi-objective ingredient optimization model, and obtain the optimal ingredient scheme. The fermentation comparison module is used to obtain the fermentation prediction results corresponding to the optimal ingredient scheme, feed raw materials according to the optimal ingredient scheme, collect real-time fermentation data during the fermentation process, and compare it with the fermentation prediction results to obtain the fermentation deviation results. The incremental update module is used to record the optimal ingredient scheme and real-time fermentation data when the fermentation deviation results exceed the preset deviation range. The optimal ingredient scheme, real-time fermentation data and finished product quality data are used as new samples, and the fermentation synergy matrix is ​​updated using an incremental learning method. The knowledge recording module is used to associate finished product quality data, optimal ingredient scheme, updated fermentation synergy matrix and raw material batch information to form a structured ingredient knowledge record; The graph construction module is used to construct an ingredient knowledge graph based on all ingredient knowledge records when the preset number of ingredient knowledge records are reached. It analyzes the deep correlation characteristics between ingredients and finished product quality, and uses the deep correlation characteristics as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of fermentation synergy effect matrix.

[0056] Through the collaborative efforts of the aforementioned components, the feature extraction module integrates the bioactivity characteristics of raw materials with the dynamic equilibrium coefficient of the carbon-nitrogen ratio into a unified representation, providing standardized raw material information input for subsequent modules. Based on this, the effect modeling module combines historical data with deep correlation features mined from the previous round of knowledge graph analysis to quantify the fermentation synergistic effects among raw materials and construct a matrix. The optimization solution module uses the comprehensive feature vector and the synergistic effect matrix as input to generate the optimal ingredient formulation through multi-objective optimization. The fermentation comparison module executes the ingredient formulation and collects real-time fermentation data, comparing it with the predicted results to generate deviation information. The incremental update module uses the deviation as a trigger condition to inject new samples into the synergistic effect matrix for local correction. The knowledge recording module archives the data of each batch of ingredients, fermentation, and finished product, accumulating structured materials for knowledge graph construction. After the data accumulation reaches a threshold, the graph construction module extracts deep correlation features and sends them back to the effect modeling module to participate in a new round of synergistic effect calculation. Each module forms a progressive closed loop from raw material perception to solution generation, from fermentation feedback to matrix update, and from knowledge accumulation to prior knowledge injection. This allows the ingredient optimization process to continuously converge as production batches progress, effectively improving the activity consistency, nutrient compliance rate, and raw material utilization efficiency of the finished enzyme fertilizer.

[0057] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the intelligent batching optimization method for enzyme fertilizer.

[0058] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligent ingredient optimization of enzyme fertilizer, characterized in that, include: S1. Collect batch information and characteristic parameters of each batch of raw materials in the batching equipment, extract bioactivity characteristics, calculate the dynamic balance coefficient of carbon-nitrogen ratio, and fuse them to obtain the comprehensive feature vector of each batch of raw materials. S2. Obtain historical ingredient plans and corresponding fermented product data to establish a standardized dataset. Combined with the deep correlation features generated in the previous round, calculate the fermentation synergy coefficient of different raw material combinations and construct a fermentation synergy matrix. S3. Construct a multi-objective ingredient optimization model based on the comprehensive feature vector and the fermentation synergy effect matrix, solve the multi-objective ingredient optimization model, and obtain the optimal ingredient scheme; S4. Obtain the fermentation prediction result corresponding to the optimal ingredient scheme, feed the raw materials according to the optimal ingredient scheme, collect real-time fermentation data during the fermentation process, and compare it with the fermentation prediction result to obtain the fermentation deviation result. S5. When the fermentation deviation results exceed the preset deviation range, record the optimal ingredient scheme and real-time fermentation data. Use the optimal ingredient scheme, real-time fermentation data and finished product quality data as new samples, and use the incremental learning method to update the fermentation synergy matrix. S6. Link the finished product quality data, the optimal ingredient scheme, the updated fermentation synergy matrix, and the batch information of raw materials to form a structured ingredient knowledge record; S7. When the ingredient knowledge records reach the preset number, construct an ingredient knowledge graph based on all ingredient knowledge records, analyze the deep correlation characteristics between ingredients and finished product quality, and use the deep correlation characteristics as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of fermentation synergy effect matrix.

2. The method according to claim 1, characterized in that, S1 includes: The raw materials in each batch of the batching equipment are divided into organic waste raw materials, enzyme strain raw materials and auxiliary materials. The characteristic parameters corresponding to each type of raw material are extracted and integrated to form a set of characteristic parameters. The bioactivity characteristics of various raw materials are extracted from the set of characteristic parameters. The bioactivity characteristics include the microbial activity index and microbial community diversity index corresponding to each raw material. Based on the current carbon-nitrogen ratio of each raw material and the preset target optimal carbon-nitrogen ratio for fermentation, the dynamic balance coefficient of the carbon-nitrogen ratio of each raw material is calculated. The set of characteristic parameters, bioactivity characteristics and dynamic balance coefficient of carbon-nitrogen ratio are integrated to construct a comprehensive feature vector for each raw material. The comprehensive feature vector includes four feature subspaces: nutrient dimension, enzyme activity dimension, microbial activity dimension and nutrient balance dimension.

3. The method according to claim 1, characterized in that, S2 includes: Historical ingredient plans and corresponding fermented product data were extracted from the historical production database. Abnormal batch data were removed and standardized to obtain a standardized dataset. Based on the standardized dataset and combined with the deep association features generated in the previous round, all raw materials were combined in pairs. The ratio of the two raw materials in each combination was used as the independent variable and the enzyme activity of the finished product was used as the dependent variable. The nonlinear response surface was obtained by fitting the local weighted regression scatter smoothing method. Based on the nonlinear response surface, the peak value of enzyme activity under the nutrient constraints of the finished product and its corresponding ratio range are determined. The predicted value of enzyme activity when each raw material is used alone is calculated. The ratio of the peak value of enzyme activity to the predicted value of enzyme activity is determined as the fermentation synergy coefficient of the corresponding raw material combination, and all fermentation synergy coefficients are obtained. All fermentation synergy coefficients are symmetricized and sparsely constrained to form a fermentation synergy matrix.

4. The method according to claim 1, characterized in that, S3 includes: Based on the comprehensive feature vector and fermentation synergy matrix, three optimization objectives are set: maximizing enzyme activity, minimizing raw material cost, and minimizing fermentation cycle. At the same time, hard constraints on finished product nutrient indicators and soft constraints on carbon-nitrogen ratio balance are set to construct a multi-objective ingredient optimization model. A pre-defined non-dominated sorting genetic algorithm is used to solve the multi-objective batching optimization model. Population screening is performed through fast non-dominated sorting and crowding calculation. After iterative convergence, the Pareto optimal solution set is obtained. The optimal ingredient ratio is selected from the Pareto optimal solution set, based on the preset decision rules.

5. The method according to claim 1, characterized in that, S4 includes: Based on the optimal ingredient ratio, the feeding control instructions for each raw material are generated, and the ingredient feeding execution mechanism is controlled to complete the feeding of raw materials and start the fermentation process. A fermentation process prediction model is established based on the optimal ingredient scheme to predict the changes in temperature, pH value, dissolved oxygen concentration, carbon dioxide release rate and enzyme activity parameters during fermentation, and to generate fermentation prediction results. During the fermentation process, the temperature, pH value, dissolved oxygen concentration, carbon dioxide release rate, and enzyme activity parameters inside the fermenter are collected to obtain real-time fermentation data. The real-time fermentation data is compared one by one with the corresponding parameters in the fermentation prediction results, the deviation values ​​of each parameter are calculated, and the results are integrated to form the fermentation deviation results.

6. The method according to claim 1, characterized in that, S5 includes: The fermentation deviation results are compared with the preset deviation range to determine whether the fermentation deviation results exceed the preset deviation range. If so, the sample collection and matrix update process is started to record the optimal ingredient scheme and real-time fermentation data. After the corresponding batch of fermentation is completed, the finished product quality data is obtained, and the optimal ingredient scheme, real-time fermentation data and finished product quality data after fermentation are determined as new samples. An incremental learning method is used to locally correct the raw material combination coefficients of the deviation correlation in the fermentation synergy matrix based on the new samples, so as to obtain the updated fermentation synergy matrix.

7. The method according to claim 1, characterized in that, S6 includes: Based on the finished product quality data corresponding to the current fermentation batch, the optimal ingredient scheme, the updated fermentation synergy matrix, and the batch information of the raw materials, determine the multi-source data to be associated; Based on the multi-source data to be associated, the batch information of raw materials is used as a unified association identifier. The finished product quality data, the optimal ingredient scheme and the updated fermentation synergy matrix are matched and bound with the corresponding batch information one by one. The matched full data is then structured and regularized to generate the corresponding batch of structured ingredient knowledge records.

8. The method according to claim 1, characterized in that, S7 includes: Count the total number of generated ingredient knowledge records. When the total number reaches the preset number, retrieve all ingredient knowledge records as the data source for graph construction. Based on the graph, a data source is constructed, and the correspondence between the ingredient parameters and the finished product quality is extracted from each ingredient knowledge record to form an ingredient knowledge graph. By performing association mining analysis on the ingredient knowledge graph, we can extract the deep correlation features between ingredients and finished product quality. Deeply correlated features are used as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of the fermentation synergy effect matrix.

9. An intelligent batching optimization system for enzyme fertilizer, used to implement the method as described in any one of claims 1-8, characterized in that, include: The feature extraction module is used to collect batch information and feature parameters of each batch of raw materials in the batching equipment, extract bioactivity features, calculate the dynamic balance coefficient of carbon-nitrogen ratio, and fuse them to obtain the comprehensive feature vector of each batch of raw materials. The effect modeling module is used to obtain historical ingredient schemes and corresponding fermented product data to establish a standardized dataset, and combine it with the deep correlation features generated in the previous round to calculate the fermentation synergy effect coefficients of different raw material combinations and construct a fermentation synergy effect matrix. The optimization and solution module is used to construct a multi-objective ingredient optimization model based on the comprehensive feature vector and the fermentation synergy effect matrix, solve the multi-objective ingredient optimization model, and obtain the optimal ingredient scheme. The fermentation comparison module is used to obtain the fermentation prediction results corresponding to the optimal ingredient scheme, feed raw materials according to the optimal ingredient scheme, collect real-time fermentation data during the fermentation process, and compare it with the fermentation prediction results to obtain the fermentation deviation results. The incremental update module is used to record the optimal ingredient scheme and real-time fermentation data when the fermentation deviation results exceed the preset deviation range. The optimal ingredient scheme, real-time fermentation data and finished product quality data are used as new samples, and the fermentation synergy matrix is ​​updated using an incremental learning method. The knowledge recording module is used to associate finished product quality data, optimal ingredient scheme, updated fermentation synergy matrix and raw material batch information to form a structured ingredient knowledge record; The graph construction module is used to construct an ingredient knowledge graph based on all ingredient knowledge records when the preset number of ingredient knowledge records are reached. It analyzes the deep correlation characteristics between ingredients and finished product quality, and uses the deep correlation characteristics as prior data to participate in the construction of the next round of multi-objective ingredient optimization model and the generation of fermentation synergy effect matrix.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the intelligent batching optimization method for enzyme fertilizer as described in any one of claims 1-8.

Images